Generation job failures
When extraction or authoring fails, and what to do about each case.
Triage in order
- 1Read the staged progress view
It names the phase that failed rather than reporting a generic error.
- 2Retry once
Most failures are transient — a provider timeout or a rate limit. A retry resumes rather than restarting from scratch.
- 3Check whether it is one gate or all of them
One gate failing repeatedly is a content problem. Every gate failing is a platform problem.
- 4Look at the source document
Thin or badly OCR-ed source material is the most common cause of repeated failure on a single gate.
Specific cases
The course is stuck in processing.
Check the progress view. A genuinely failed job moves the course to failed and names the phase; a running job shows progress. Jobs resume rather than restart when a worker is replaced, so a deploy mid-run is not the cause.
Extraction finished but the gates are wrong.
That is what Review Extracted is for — everything is editable and edits cascade. If the structure is fundamentally wrong the source document probably lacked sequence or outcomes; re-upload a fuller version rather than hand-correcting fifty rows.
One gate fails generation every time.
Its source material is likely too thin to ground generation, and generation fails rather than inventing content. Enrich that part of the syllabus, or accept the gate with less generated material.
Generation is far slower than usual.
A full course is roughly a quarter of an hour. Substantially longer suggests queue depth. It will complete; jobs are durable.
I regenerated and lost my edits.
Regeneration replaces the artefact. Edits to a lesson plan or script are yours and persist — but regenerating that artefact overwrites them. Edit after regenerating, not before.